基于傅立叶变换的人工神经网络近红外光谱定量分析法

摘要

After some pretreatments to the original spectra and using the frontal TV coefficients of the fast Fourier transform (FFT) as the input variables of the artificial neural network (ANN) , a lot of useful information was assured to participate in modeling, and the advanced filter function of the FFT was also realized. After modeling the octane in gasoline and the calorific value in coal powder, the FFT - RBF(the radial basis function network) model was found to be good, for example, when using the frontal 20 coefficients of FFT, the root mean square error(RMSEP) of prediction of the octane is 0. 152, and its correlation coefficient is 0. 976, and when using the frontal 30 coefficients of FFT, the RMSEP of the Qgr. D of the coal powder is 0. 256, and its correlation coefficient is 0. 923 . The research illustrated that the ANN NIR quantitative analysis method based on the FFT, especially the FFT - RBF had the tremendous advantage in NIR prediction function.%将原始光谱进行一定的预处理后,以其快速傅立叶变换FFT的前N个系数作为人工神经网络(ANN)的输入量,不仅确保了大量有用信息参与模型的建立,同时实现了优越的滤波功能.以汽油的辛烷值和煤粉干燥基高位发热量(Qgr.d)的近红外光谱建模,当采用前20个FFT系数的傅立叶变换-径向基网络( FFT - RBF)时,辛烷值模型的预测误差均方根(RMSEP)可达0.152,相关系数为0.976,当采用前30个FFT系数时,快速FFT - RBF煤粉干燥基高位发热量模型的RMSEP为0.256,相关系数为0.923,说明FFT -RBF模型有着很好的预测能力.研究表明基于傅立叶变换的人工神经网络近红外光谱定量分析法,特别是FFT - RBF具有良好的预测能力.